A seven-step workflow you can reuse
If you want the practical sequence before the case study, use these seven steps:
- Prepare the source packet. Gather approved copy, data, brand rules, the target slide count, and the office application the recipient will use.
- Write a title-only outline. Give every slide one job. Remove pages that repeat the theme without moving the story forward.
- Define a small design system. Choose the type hierarchy, colors, margins, image treatment, and three to five layout families before building individual pages.
- Name the objects that must remain editable. Text, charts, tables, timelines, and process diagrams usually need to survive handoff as real presentation objects.
- Create imagery for assigned layouts. Decide the crop, subject position, lighting, and copy space before generating or selecting each image.
- Build, render, and repair. Export a draft, render every slide, inspect the contact sheet and difficult pages at full size, then fix the source rather than covering defects in the export.
- Test the handoff. Open the deck in the recipient’s application, change representative text and data, save it, reopen it, and record what you did not test.
This sequence takes longer than asking for “a beautiful presentation,” but it leaves fewer hidden problems for the person who has to edit the deck later.
Where this workflow fits
This article focuses on a premium, Codex-specific build where visual quality and editability matter equally. If you need a broader introduction to outlines, source files, notes, accessibility, and final sharing, start with How to Create an Editable PowerPoint With ChatGPT. If you are still choosing a tool, the AI presentation-maker comparison covers free and trial export routes. If your starting point is an existing deck that needs to become a documented report, use the PowerPoint-to-report workflow.
The route below is most useful when another person must continue editing the file, when the deck needs an art-directed visual system, or when charts and tables cannot be flattened. It is probably more work than you need for a disposable internal update or a presentation that will only be viewed as a hosted link.
What we rebuilt and what we left behind
The starting reference was a short X video in which a user asks a presentation Skill to make an Apple Watch and sports deck. The result uses green, cream, and black, strong photography, oversized type, and varied compositions. It is a useful quality benchmark, but it is not a production specification.
We did not install or extract the private instructions of the demonstrated Skill. We also did not reuse Apple trademarks, product photographs, slide copy, or layouts. We created a new concept called Aster Move, generated original unbranded visuals, wrote new copy, and designed an independent eight-slide story. The product and all numerical values are illustrative so the design exercise cannot be mistaken for real product research or health advice.
The narrower question was more useful than trying to clone the reference: can a normal Codex workspace turn a clear brief into a presentation that feels art-directed and still hands over cleanly?
Decide what “the same effect” means
People use “AI PowerPoint” for three different deliverables:
| Deliverable | What looks finished | What the recipient can change | Main risk |
|---|---|---|---|
| Browser presentation | A hosted interactive page | Whatever the web editor exposes | No dependable PPTX handoff |
| Slide images inside PPTX | Every page can look exact | Usually only whole-slide images | The extension says PowerPoint, but the content is flattened |
| Editable PPTX | A normal presentation in an office app | Text, charts, tables, shapes, and notes | More engineering and more compatibility testing |
We targeted the third row. The chart needed real chart data and an embedded workbook, not a picture of bars. The four-week plan needed individual table cells. Visual polish still mattered, but editability became an acceptance rule instead of a feature we hoped would survive.
The brief that kept the project honest
A vague prompt would make it impossible to tell whether the output had solved the assignment or merely looked attractive. We used a compact brief with a real audience, one emotional direction, named slide jobs, and explicit handoff rules:
Create an eight-slide, 16:9 editorial concept deck for Aster Move,
a fictional wearable idea for people returning to exercise.
Audience: a creative review team.
Tone: calm, premium, human, and quietly technical.
Visual system: dark forest green, warm ivory, acid-lime accents,
large editorial typography, restrained rules, generous whitespace.
Required content:
1. Cover
2. Human premise
3. Editable seven-day effort chart
4. Editable three-step process diagram
5. Editable four-week plan table
6. Recovery interlude
7. Design-system synthesis
8. Minimal close
Boundaries:
- The product, people, figures, and performance claims are fictional.
- Do not use Apple branding or copied campaign material.
- Keep text, chart, table, and process objects editable.
- Put limitations and fictional-data notices in speaker notes.
- Render every slide and repair visible defects before delivery.
The brief does not prescribe the exact location of every box. It fixes the audience, tone, visual grammar, page responsibilities, editing requirements, and claim limits. Codex still has room to design, but it cannot quietly change the deliverable.
Turn the method into a reusable Codex Skill
OpenAI’s current Skill documentation describes a Skill as a folder with a required SKILL.md and optional scripts, references, assets, and interface metadata. A Skill can be invoked directly or selected when its description matches the task. For this project, that structure is more useful than one oversized “make it premium” prompt because it preserves the decisions that should survive across different presentation topics.
The downloadable Grow AI Skills PPT Skill has three working layers:
- Workflow. Inspect the sources, agree on the outline, define one job per slide, build the editable objects, render the deck, repair visible defects, and deliver the evidence with the file.
- Design guidance. Use a restrained palette, large type, generous margins, a small set of layout families, and images composed for their assigned crop.
- Quality gates. Reject flattened evidence, clipped text, accidental overlap, unsupported claims, missing notes, and exports that nobody has inspected.
The Skill does not force every presentation to use Aster Move’s colors. It preserves the reviewable production method. For a new project, replace the design tokens and brief while keeping the source and quality checks.
Use the outline to control pacing
A common AI deck failure is eight versions of the same card: heading, three bullets, decorative image. We prevented that before opening the layout tool by reading the titles as a sequence:
- Movement, with less noise
- People return to exercise in uneven weeks
- Weekly effort, made readable
- Three steps shape the daily cue
- Progress stays visible without crowding the slide
- Recovery belongs in the same story as effort
- One visual language, three moments
- Move enough to notice
The order moves from an emotional opening into the premise, the system, a plan, a visual pause, a design summary, and a short close. Only after that did we assign the layout families: split-image cover, photo-and-copy premise, chart page, process diagram, table, full-bleed interlude, typographic summary, and minimal close.
Generate images for their assigned space
We generated four separate assets: an athlete fastening an unbranded watch, a quiet city return-to-motion scene, a training-kit flat lay, and a recovery scene. None contains baked-in titles, logos, buttons, or fake interface text.
The cover needed shadow and open space on the left. The bridge image needed its subject offset so the slide could reserve a copy column. The flat lay had to survive a narrow crop. The recovery scene needed enough surrounding space to slow the presentation down after three information-heavy pages.
A request for “a beautiful fitness image” often returns a centered subject with nowhere to place text. Specify the crop, subject position, lighting, surface, and forbidden text before generation. Keep the actual presentation copy editable on the slide.

Build the evidence slides for future edits
The chart
Slide three uses seven illustrative effort values: 22, 38, 58, 28, 44, 72, and 16. The final package contains an editable bar chart, seven cached values, and an embedded workbook. In Keynote, we could select the chart and open its data editor to see the seven labels and values.

The low, steady, and push labels on the right are separate text objects. They interpret the values while the chart owns the data. An editor can change the numbers without rebuilding the entire page.
The table
Slide five uses a five-row, four-column table for the illustrative four-week plan. The photograph on the right is an image crop; the title and explanation remain ordinary text boxes. Keynote exposed the plan as individual cells rather than one flattened picture.

The practical question is simple: can the recipient change “3 × 25 min” to “2 × 25 min” without opening an image editor or covering old pixels with a new box? If not, the slide may look finished, but it is a poor handoff.
The failed first pass was more useful than a perfect demo
The first export was structurally healthy, but visual review found three problems. On slide three, explanatory copy competed with the chart. On slide four, the right-side statement crowded the last process step. On slide seven, the two-line title dropped into the supporting sentence.

The slide-seven comparison makes the problem obvious. In the rejected version at the top, the title wraps and overlaps the sentence below it. In the repaired version at the bottom, the shorter title fits on one line and the spacing creates a clear reading order. We also adjusted widths and spacing on slides three and four, rebuilt the file, and rendered all eight pages again.
The second pass returned zero findings and zero warnings from the layout validator. That did not make the design automatically good. It confirmed that the repaired geometry no longer triggered the specific checks, while the new renders gave us something visual to judge. OpenAI’s slide-deck guidance makes the same practical point: preview and review the final presentation before delivery.
What the checks support
| Check | Result | What we learned | Remaining limit |
|---|---|---|---|
| Package integrity | Pass; 8 slides, 1 chart, 0 findings | The retained PPTX is structurally readable | Structure alone cannot judge taste |
| Layout validation | 0 findings, 0 warnings | The repaired file triggered no geometry or approved-font issue | Other office versions may render differently |
| Editable chart | Pass on slide 3 | The chart, values, and embedded workbook exist | The illustrative numbers are not research |
| Editable table | One table on slide 5 | The plan uses cells rather than a screenshot | The plan is not training advice |
| Keynote | Opened; all 8 slides reviewed | Images, text, chart data, and table cells remained usable | This is not a Microsoft PowerPoint result |
| LibreOffice | PDF export and full render passed | A second renderer preserved the intended appearance | It cannot guarantee fidelity everywhere |
| Microsoft PowerPoint | Not installed | We make no native PowerPoint application claim | Open, edit, save, slideshow, and accessibility checks remain untested there |
Install and reuse the Skill in Codex
For a personal local installation, unzip the download so the main file is located at ~/.agents/skills/growaiskills-ppt/SKILL.md. Codex scans that user-level skills directory and normally detects changes automatically. If the Skill does not appear, restart Codex. You can then mention $growaiskills-ppt directly in your request, or let Codex select it when the task matches its description.
Before starting a real deck, prepare six inputs:
- the approved source material and any facts that remain unresolved;
- the audience and the decision or message the deck must support;
- the slide count and presentation format;
- brand rules, a template, or a reference deck if one must be preserved;
- the objects that another person must be able to edit;
- the target application and platform for the final handoff.
A useful first request is:
Use $growaiskills-ppt to create an editable 16:9 PowerPoint.
Before building:
1. Inspect my source files and list unresolved facts.
2. Propose a title-only outline with one job per slide.
3. Define a compact design system and three to five layout families.
4. Tell me which objects must remain editable.
During the build:
- Keep claims inside the supplied evidence.
- Generate or source images for their assigned crop and copy space.
- Use editable charts, tables, shapes, and text when future changes matter.
- Add speaker notes for source context and limitations.
Before delivery:
- Render every slide.
- Inspect a contact sheet and the difficult slides at full size.
- Repair clipping, overlap, weak contrast, and repetitive composition.
- Validate the PPTX package and report which office app was opened.
- Deliver the PPTX, build source, contact sheet, and review record.
If you have no factual source material, ask for a concept deck and label illustrative data. “Make it premium” should never become permission to invent market figures, product capabilities, user research, or customer quotes.
The final PowerPoint handoff still needs a human check
For a high-stakes client presentation, the person delivering the file should complete a final pass in the recipient’s version of PowerPoint:
- change one title, one chart value, and one table cell;
- save, close, and reopen the file;
- confirm that approved fonts have not been substituted;
- run the Accessibility Checker and review reading order and alt text;
- test slideshow mode, links, video, and presenter notes where applicable;
- open the exact sharing copy and check for hidden slides, comments, or unintended metadata.
Microsoft’s PowerPoint accessibility guidance is a useful reference for that last review. Our retained project did not complete the native PowerPoint portion because the application was not installed on the test machine.
Does this avoid paying for Skillry?
This route avoids buying or depending on the particular third-party presentation Skill shown in the reference. It uses Codex’s presentation capability, a local reusable Skill, original images, and local validation. The workflow files remain inspectable and reusable.
That does not make the process free in every sense. You still need access to Codex, and image or model usage may count against the limits of your plan. You also spend time reviewing the content, repairing the layout, and testing the handoff. For a simple internal deck, approved photographs or existing icons may be enough. For an important external presentation, the application check and rehearsal are part of the cost of doing the work properly.
Download the project files
- Download the reviewed editable PPTX with eight slides, original visuals, an editable chart, a real table, diagram shapes, and speaker notes.
- Download the Grow AI Skills PPT Skill with the workflow, design guidance, and quality gates.
- Download the complete review kit with the deck, generated assets, builder, Skill, validation receipt, contact sheet, and review record.
- Download the test record for the file hashes, application results, iteration history, and stated limits.
What we would repeat on the next deck
The visual quality did not come from a secret sentence in the prompt. It came from a short brief, a title sequence that controlled the pace, images made for specific crops, and editable objects chosen before layout work began.
The part we would keep most carefully is the rejected first pass. It exposed problems that the file checks could not judge and forced the team to make three small, specific repairs. That is a more dependable way to produce a premium presentation than expecting one prompt or one export to be the finished answer.
